README.md

---
license: cc-by-4.0
library_name: ultralytics
pipeline_tag: object-detection
tags:
- wildlife
- yolo
- yolo26
- object-detection
- camera-trap
- bobcat
---

# Model Card — Bobcat (*Lynx rufus*)

Single-class detection model for Bobcat, fine-tuned from the Ultralytics
YOLO26s backbone (pretrained on COCO).

**Model file:** `yolo26s_finetuned_bobcat_by_J.Gong_uwyo_2026-05-28.pt`

## Training Details

| Property | Value |
|----------|-------|
| Base model | yolo26s.pt (COCO pretrained, Ultralytics) |
| Architecture | YOLO26s |
| Input size | 640 × 640 |
| Epochs | 150 |
| Optimizer | MuSGD, lr=0.002, momentum=0.9 |
| Augmentation | mosaic=1.0, degrees=10°, scale=0.5, fliplr=0.5, hsv_h/s/v |
| Device | NVIDIA RTX 5000 Ada Generation (32 GB, CUDA 12.8) |
| Training date | 2026-05-28 |
| Author | Jian Gong, University of Wyoming |

## Dataset

Images sourced from iNaturalist (research-grade observations).
Bounding boxes generated by MegaDetector v5a (confidence ≥ 0.15).
Split 80 / 10 / 10 train / val / test.

| Split | Images |
|-------|-------:|
| train | 189 |
| val   | 23 |
| test  | 25 |

## Performance

Evaluated on the held-out validation set (best checkpoint).

| Metric | Value |
|--------|------:|
| mAP50 | 0.6649 |
| mAP50-95 | 0.5188 |

## Usage

```python
from ultralytics import YOLO
model = YOLO("models/bobcat/yolo26s_finetuned_bobcat_by_J.Gong_uwyo_2026-05-28.pt")
results = model.predict("image.jpg", conf=0.25)
```